WB Project PDO text EDA

Work in progress

TO DO [COMPL🟠🟡] list

  • check if the peaks in sector words correspond to peak in sector (the feature)
  • add vertical lines with relevant WDR over the years
  • do bigram plots on “climate change” and “climate resilient”

Set up

# Pckgs -------------------------------------
library(fs) # Cross-Platform File System Operations Based on 'libuv'
library(here) # A Simpler Way to Find Your Files
library(paint) # paint data.frames summaries in colour
library(tidyverse) # Easily Install and Load the 'Tidyverse' 
library(janitor) # Simple Tools for Examining and Cleaning Dirty Data
library(skimr) # Compact and Flexible Summaries of Data
library(readxl) # Read Excel Files
library(kableExtra) # Construct Complex Table with 'kable' and Pipe Syntax)
library(tidytext) # Text Mining using 'dplyr', 'ggplot2', and Other Tidy Tools
library(patchwork) # The Composer of Plots

# # ML ------------------------------------- 
# library(tidymodels) # Easily Install and Load the 'Tidymodels' Packages 
# library(textrecipes) # Extra 'Recipes' for Text Processing 

# # TEXT ANALYTICS -------------------------------------
# library(tidytext) # Text Mining using 'dplyr', 'ggplot2', and Other Tidy Tools 
# library(SnowballC) # Snowball Stemmers Based on the C 'libstemmer' UTF-8 Library
# library(cleanNLP) # A Tidy Data Model for Natural Language Processing

—————————————————————————-

Data sources

The data used in this analysis comes from the World Bank’s Projects & Operations database.

Since some pre-processing steps are computationally expensive, I did that in a separate step (analysis/_01a_WB_project_pdo_prep.qmd), WHERE:

  1. Retrieved manually ALL WB projects (22,569) approved between FY 1947 and 2026 as of 31/08/2024 using simply the Excel button on this page WBG Projects

    • Of these, approximately 50% (11,322 projects) had a “viable” PDO text in the dataset (i.e., not blank or labeled as “TBD”, etc.).
    • There are no Project Development Objectives available in projects approved before FY2001
    • However, other than approval year, based on some tests on available projects’ features, PDO texts’ missingness seems to happen at random.

  2. Dropped from the analysis: projects with no PDO text (including those approved before FY2001), and projects with missing status, FY of approval –> 8,811 usable projects selected.

  3. Split the dataset into training / validation / test subsets (proportional to FY and regional distribution).

    • Here I work only on training set (~50% of usable ones, i.e. 4,403 PDOs).

  4. Cleaned the data (parse dates, recode variables, etc.), fix typos, unwanted special characters, and other unimportant issues in PDOs.

  5. Obtained PoS tagging + tokenization with cleanNLP package (functions cnlp_init_udpipe() + cnlp_annotate()) and saved projs_train_t (cleaned train dataset).

cleanNLP supports multiple backends for processing text, such as CoreNLP, spaCy, udpipe, and stanza. Each of these backends has different capabilities and might require different initialization procedures.

  • CoreNLP ~ powerful Java-based NLP toolkit developed by Stanford, which includes many linguistic tools like tokenization, part-of-speech tagging, and named entity recognition.
    • ❕❗️ NEEDS EXTERNAL INSTALLATION (must be installed in Java with cnlp_install_corenlp() which installs the Java JAR files and models)
  • spaCy ~ fast and modern NLP library written in Python. It provides advanced features like dependency parsing, named entity recognition, and tokenization.
    • ❕❗️ NEEDS EXTERNAL INSTALLATION (fmst be installed in Python (with spacy_install() which installs both spaCy and necessary Python dependencies) and the spacyr R package must be installed to interface with it.
  • udpipe ~ R package that provides bindings to the UDPipe NLP toolkit. Fast, lightweight and language-agnostic NLP library for tokenization, part-of-speech tagging, lemmatization, and dependency parsing.
  • stanza~ another modern NLP library from Stanford, similar to CoreNLP but built on PyTorch and supports over 66 languages…

when you initialize a back-end (like CoreNLP) in cleanNLP, it stays active for the entire session unless you reinitialize or explicitly change it.

# ---- 1) Initialize the CoreNLP backend
library(cleanNLP) # A Tidy Data Model for Natural Language Processing # A Tidy Data Model for Natural Language Processing # A Tidy Data Model for Natural Language Processing
cnlp_init_corenlp()
# If you want to specify a language or model path:
cnlp_init_corenlp(language = "en", 
                  # model_path = "/path/to/corenlp-models"
                  )

# ---- 2) Initialize the spaCy backend 
library(cleanNLP) # A Tidy Data Model for Natural Language Processing # A Tidy Data Model for Natural Language Processing # A Tidy Data Model for Natural Language Processing
library(spacyr) # Wrapper to the 'spaCy' 'NLP' Library
# Initialize spaCy in cleanNLP
cnlp_init_spacy()
# Optional: specify language model
cnlp_init_spacy(model_name = "en_core_web_sm")

# ---- 3) Initialize the udpipe backend
library(cleanNLP) # A Tidy Data Model for Natural Language Processing # A Tidy Data Model for Natural Language Processing # A Tidy Data Model for Natural Language Processing
# Initialize udpipe backend
cnlp_init_udpipe(model_name = "english")

# ---- 4) Initialize the stanza backend

[TBL] Illustrative PDOs text in Projects’ documents

Project_ID Project_Name Project_Development_Objective
P127665 Second Economic Recovery Development Policy Loan This development policy loan supports the Government of Croatia's reform efforts with the aim to: (i) enhance fiscal sustainability through expenditure-based consolidation; and (ii) strengthen investment climate.
P069934 PERNAMBUCO INTEGRATED DEVELOPMENT: EDUCATION QUALITY IMPROVEMENT PROJECT The development objectives of the Pernambuco Integrated Development: Education Quality Improvement Project are to (a) improve the quality, efficiency, and inclusiveness of the public education system; (b) modernize and strengthen the managerial, financial, and administrative capacity of the Secretariat of Education to set policies and guidelines for the sector and deliver public education efficiently; and (c) support the overall state modernization effort through interventions to be carried out in the Secretariat of Education and to be replicated in other state institutions.

Notes on PDO text data quality

First, it is important to notice that all 7,548 projects approved before FY2001 had no PDO text available.

The exploratory analysis of the 11,353 projects WITH PDO text revealed some interesting findings:

  1. PDO text length: The PDO text is quite short, with a median of 2 sentences and a maximum of 9 sentences.
  2. PDO text missingness: besides 11,306 projects with missing PDOs, 31 projects had some invalid PDO values, namely:
    • 11 have PDO as one of: “.”,“-”,“NA”, “N/A”
    • 7 have PDO as one of: “No change”, “No change to PDO following restructuring.”,“PDO remains the same.”
    • 9 have PDO as one of: “TBD”, “TBD.”, “Objective to be Determined.”
    • 4 have PDO as one of: “XXXXXX”, “XXXXX”, “XXXX”, “a”

Of the remaining 11,322 projects with a valid PDO, some more projects were excluded from the analysis for incompleteness:

  • 3 projects without “project status”
  • 2,176 projects without “board approval FY”
  • 332 projects approved in FY >= FY2024 (for incomplete approval stage)

Lastly (and this was quite surprising to me) the remaining, viable 8,811 unique projects, were matched by only 7,582 unique PDOs! In fact, 2,235 projects share 1,006 NON-UNIQUE PDO text in the “cleaned” dataset. Why? Apparently, the same PDO is re-used for multiple projects (from 2 to as many as 9 times), likely in cases of follow-up phases of a parent project or components of the same lending program.”

In sum, the cleaning process yielded a usable set of 8,811 functional projects, which was split into a training subset (4,403) to explore and test models and a testing subset (4408), held out for post-prediction evaluation.

Note

Evidently, in some cases,the same PDO is used for multiple projects (from a minimum of 2 to a maximum of 9 time!!!), most likely when there is a parent project or subsequent phases of the same lending program.

—————————————————————————

Load pre-processed Projs’ dataset + PDO dataset

Here I will just load the pre-processed data (training set only).

[Saved file projs_train_t & pdo_train_t]

# Load Proj train dataset `projs_train_t`
projs_train <- readRDS("~/Github/slogan/data/derived_data/projs_train.rds")  

# Load clean tokenized-PDO dataset `pdo_train_t`
pdo_train_t <- readRDS(here::here("data" , "derived_data", "pdo_train_t.rds"))

Previous Tokenization and PoS Tagging

Typically, one of the first steps in this transformation from natural language to feature, or any of kind of text analysis, is tokenization.

i) Explain Tokenization

Breaking units of language into components relevant for the research question is called “tokenization”. Components can be words, n-grams, sentences, etc. or combining smaller units into larger units.

  • Tokenization is a row-wise operation: it changes the number of rows in the dataset.

The choices of tokenization

  1. Should words be lower cased?
  2. Should punctuation be removed?
  3. Should numbers be replaced by some placeholder?
  4. Should words be stemmed (also called lemmatization)? ☑️
  5. Should bigrams/multi-word phrase be used instead of single word phrases? ☑️
  6. Should stopwords (the most common words) be removed? ☑️
  7. Should rare words be removed? ❌
  8. Should hyphenated words be split into two words? ❌

for the moment I keep all as conservatively as possible

ii) Explain Pos Tagging

Linguistic annotation is a common for of enriching text data, i.e. adding information about the text that is not directly present in the text itself.

Upon this, e.g. classifying noun, verb, adjective, etc., one can discover intent or action in a sentence, or scanning “verb-noun” patterns.

Here I have a training dataset file with:

Variable Type Provenance Description Example
proj_id chr original PDO data
pdo chr original PDO data
word chr original PDO data Governments
sid int output cleanNLP sentence ID
tid chr output cleanNLP token ID within sentence
token chr output cleanNLP Tokenized form of the token. government
token_with_ws chr output cleanNLP Token with trailing whitespace government
lemma chr output cleanNLP The base form of the token government
stem chr output SnowballC The base form of the token govern
upos chr output cleanNLP Universal part-of-speech tag (e.g., NOUN, VERB, ADJ).
xpos chr output cleanNLP Language-specific part-of-speech tags.
feats chr output cleanNLP Morphological features of the token
tid_source chr output cleanNLP Token ID in the source document
relation chr output cleanNLP Dependency relation between the token and its head token
pr_name chr output cleanNLP Name of the parent token
FY_appr dbl original PDO data
FY_clos dbl original PDO data
status chr original PDO data
regionname chr original PDO data
countryname chr original PDO data
sector1 chr original PDO data
theme1 chr original PDO data
lendinginstr chr original PDO data
env_cat chr original PDO data
ESrisk chr original PDO data
curr_total_commitment dbl original PDO data

— PoS Tagging: upos (Universal Part-of-Speech)

upos n percent explan
ADJ 21261 0.0852623 Adjective
ADP 27050 0.1084777 Adposition
ADV 2950 0.0118303 Adverb
AUX 3588 0.0143888 Auxiliary
CCONJ 14236 0.0570902 Coordinating conjunction
DET 21505 0.0862408 Determiner
INTJ 57 0.0002286 Interjection
NOUN 70752 0.2837344 Noun
NUM 2190 0.0087825 Numeral
PART 8691 0.0348532 Particle
PRON 2330 0.0093439 Pronoun
PROPN 14856 0.0595765 Proper noun
PUNCT 28393 0.1138635 Punctuation
SCONJ 2160 0.0086622 Subordinating conjunction
SYM 316 0.0012672 Symbol
VERB 25806 0.1034889 Verb
X 3219 0.0129090 Other

On random visual check, these are not always correct, but they are a good starting point for now.

iii) Custom Stopwords

Remove stop words, which are the most common words in a language.

  • but I don’t want to remove any meaningful word for now
# Custom list of articles, prepositions, and pronouns
custom_stop_words <- c(
   # Articles
   "the", "a", "an",   
   "and", "but", "or", "yet", "so", "for", "nor", "as", "at", "by", "per",  
   # Prepositions
   "of", "in", "on", "at", "by", "with", "about", "against", "between", "into", "through", 
   "during", "before", "after", "above", "below", "to", "from", "up", "down", "under",
   "over", "again", "further", "then", "once",  
   # Pronouns
   "i", "me", "my", "myself", "we", "our", "ours", "ourselves", "you", "your",
   "yours", "yourself", "yourselves", "he", "him", "his", "himself", "she", "her", 
   "hers", "herself", "it", "its", "itself", "they", "them", "their", "theirs", "themselves" ,
   "this", "that", "these", "those", "which", "who", "whom", "whose", "what", "where",
   "when", "why", "how", "all", "any", "both", "each", "few", "more", "most", "other",
   # "some", "such", "no",  "not", 
   # "too", "very",   
   # verbs
   "is", "are", "would", "could", "will", "be", "e.g", "e.g.", "i.e.",
   "i", "ii", "iii", "iv", "v",
   # because tautology
   "pdo"
)

# Convert to a data frame if needed for consistency with tidytext
custom_stop_words_df <- tibble(word = custom_stop_words)
saveRDS(custom_stop_words, here("data" , "derived_data", "custom_stop_words.rds"))
saveRDS(custom_stop_words_df, here("data" , "derived_data", "custom_stop_words_df.rds"))

iv) Stemming

Often documents contain different versions of one base word, often called a stem. Stemming is the process of reducing words to their base or root form.

Snowball is one framework released in 1980 with an open-source license that can be found in R package SnowballC.

# Using `SnowballC::wordStem` to stem the words. e.g.
pdo_train_t <- pdo_train_t %>% 
   mutate(stem = SnowballC::wordStem(token_l)) %>%
   relocate(stem, .after = lemma)

Why Stemming?: For example, in topic modeling, stemming reduces noise by making it easier for the model to identify core topics without being distracted by grammatical variations. (Lemmatization is more computationally intensive as it requires linguistic context and dictionaries, making it slower, especially on large datasets)

Token Lemma Stem
development development develop
quality quality qualiti
high-quality high-quality high-qual
include include includ
logistics logistic logist
government/governance Governemnt/government/governance govern

NOTE: Among word / stems encountered in PDOs, there are a lot of acronyms which may refer to World Bank lingo, or local agencies, etc… Especially when looked at in low case form they don’t make much sense…

Notes on sparsity

Sparsity in the context of a document-term matrix refers to the proportion of cells in the matrix that contain zeros. High sparsity means that most terms do not appear in most documents.

  • removing stopwords before stemming can reduce sparsity
  • tidytext::cast_tdm turns a “tidy” one-term-per-document-per-row data frame into a Document-Term Matrix (DTM) from the tm package.
    • this dataset contains 4403 documents (each of them a PDO) and 11029 terms (distinct words). Notice that this DTM is 100% sparse (100% of document-word pairings are zero, bc most pairings of document and term do not occur (they have the value zero).
# create document-word matrix
DTM <- pdo_train_t %>% 
   anti_join(custom_stop_words_df, by = c("token_l" = "word")) %>% 
   count(proj_id, token_l) %>%
   tidytext::cast_dtm(proj_id, token_l, n) # HIGH!!!

DTM
# <<DocumentTermMatrix (documents: 4403, terms: 11029)>>
# Non-/sparse entries: 129940/48430747
# Sparsity           : 100%
# Maximal term length: 34
# Weighting          : term frequency (tf)

v) Document-term matrix or TF-IDF

The tf-idf is the product of the term frequency and the inverse document frequency::

\[ \begin{aligned} tf(\text{term}) &= \frac{n_{\text{term}}}{n_{\text{terms in document}}} \\ idf(\text{term}) &= \ln{\left(\frac{n_{\text{documents}}}{n_{\text{documents containing term}}}\right)} \\ tf\text{-}idf(\text{term}) &= tf(\text{term}) \times idf(\text{term}) \end{aligned} \]

— TF-IDF matrix on train pdo

# reduce size 

pdo_train_4_tf_idf <- pdo_train_t %>% # 255964
   # Keep only content words [very restrictive for now]
   # normally c("NOUN", "VERB", "ADJ", "ADV")
   filter(upos %in% c("NOUN")) %>% #    72,668 
   filter(!token_l %in% c("development", "objective", "project")) %>%   #  66,741
   # get rid of stop words (from default list)   
   filter(!token_l %in% custom_stop_words_df$word) %>%   #  66,704
   # Optional: Remove lemmas of length 1 or shorter
   filter(nchar(lemma) > 1)  #  66,350

Now, count the occurrences of each lemma for each document. (This is the term frequency or tf)

# This is the term frequency or `tf`

# Count lemmas per document
lemma_counts <- pdo_train_4_tf_idf %>%
  count(proj_id, lemma, sort = TRUE)
# Preview the result
head(lemma_counts)

With the lemma counts prepared, the bind_tf_idf() function from the tidytext package computes the TF-IDF scores.

# Compute the TF-IDF scores
lemma_tf_idf <- lemma_counts %>%
  bind_tf_idf(lemma, proj_id, n) %>%
  arrange(desc(tf_idf))

head(lemma_tf_idf)

What to use: token, lemma, or stem?

General Preference in Real-World NLP:

  • Tokens for analyses where word forms matter or for sentiment analysis.
  • Lemmas (*) for most general-purpose NLP tasks where you want to reduce dimensionality while maintaining accuracy and clarity of meaning.
  • Stems for very large datasets, search engines, and applications where speed and simplicity are more important than linguistic precision.

(*) I use lemma, after “aggressively” reducing the number of words to consider, and removing stop words (at least for now).

_______

TEXT ANALYSIS/SUMMARY

_______

We are looking at (training data subset) pdo_train_t which has 249360 rows and 26 columns obtained from 4,403 PDOs (of which 4050 unique) of 4403 Wold Bank projects approved in Fiscal Years ranging from 2001 to 2023.

[TBL] Frequencies of documents/words/stems

entity counts
N proj 4403
N PDOs 4050
N words 12953
N token 11114
N lemma 11178
N stem 8541

Term frequency

Note: normally, the most frequent words are function words (e.g. determiners, prepositions, pronouns, and auxiliary verbs), which are not very informative. Moreover, even content words (e.g. nouns, verbs, adjectives, and adverbs) can often be quite generic semantically speaking (e.g. “good” may be used for many different things).

However, in this analysis, I do not use the STOPWORD approach, but use the POS tags to reduce – in a more controlled way – the dataset, filtering the content words such as nouns, verbs, adjectives, and adverbs.

[FUNC] save plots

[FIG] Overall token freq ggplot

  • Excluding “project” “develop”,“objective”
  • Including only “content words” (NOUN, VERB, ADJ, ADV)
# Evaluate the title with glue first
title_text <- glue::glue("Most frequent TOKEN in {n_distinct(pdo_train_t$proj_id)} PDOs from projects approved between FY {min(pdo_train_t$FY_appr)}-{max(pdo_train_t$FY_appr)}") 

pdo_wrd_freq <- pdo_train_t %>%   # 123,927
   # include only content words
   filter(upos %in% c("NOUN", "VERB", "ADJ", "ADV")) %>%
   #filter (!(upos %in% c("AUX","CCONJ", "INTJ", "DET", "PART","ADP", "SCONJ", "SYM", "PART", "PUNCT"))) %>%
   filter (!(relation %in% c("nummod" ))) %>% # 173,686 
 filter (!(token_l %in% c("pdo","project", "development", "objective","objectives", "i", "ii", "iii",
                          "is"))) %>% # whne it is VERB
   count(token_l) %>% 
   filter(n > 800) %>% 
   mutate(token_l = reorder(token_l, n))   # reorder values by frequency

# plot 
pdo_wrd_freq_p <- pdo_wrd_freq %>% 
   ggplot(aes(token_l, n)) +
   geom_col(fill = "#d7b77b") +
   scale_y_continuous(breaks = seq(0, max(pdo_wrd_freq$n), by = 400)) + # directly use 'n' instead of .data$n
   coord_flip() + # flip x and y coordinates so we can read the words better
   labs(#title = title_text,
      subtitle = "[TOKEN with count > 800]", y = "", x = "")+
   geom_hline(yintercept = 800, linetype = "dashed", color = "#873c4a") +
     my_pretty_theme

[FIG] Overall stem freq ggplot

  • Without “project” “develop”,“objective”
  • Including only “content words” (NOUN, VERB, ADJ, ADV)
# Evaluate the title with glue first
title_text <- glue::glue("Most frequent STEM in {n_distinct(pdo_train_t$proj_id)} PDOs from projects approved between FY {min(pdo_train_t$FY_appr)}-{max(pdo_train_t$FY_appr)}") 
# Plot
pdo_stem_freq <- pdo_train_t %>%   # 256,632
   # include only content words
   filter(upos %in% c("NOUN", "VERB", "ADJ", "ADV")) %>%
   filter (!(relation %in% c("nummod" ))) %>% # 173,686 
   filter (!(stem %in% c("pdo", "project", "develop", "object", "i", "ii", "iii"))) %>%
   count(stem) %>% 
   filter(n > 800) %>%
   mutate(stem = reorder(stem, n))    # reorder values by frequency
   
   # plot 
pdo_stem_freq_p <-    pdo_stem_freq %>% 
      ggplot(aes(stem, n)) +
      geom_col(fill = "#d7b77b") +
      scale_y_continuous(breaks = seq(0, max(pdo_stem_freq$n), by = 400)) + # directly use 'n' instead of .data$n
      coord_flip() + # flip x and y coordinates so we can read the words better
      labs(#title = title_text,
         subtitle = "[STEM with count > 800]", y = "", x = "") +
      geom_hline(yintercept = 800, linetype = "dashed", color = "#873c4a") +
     my_pretty_theme

Evidently, after stemming, more words (or stems) reach the threshold frequency count of 800 (they have been combined by root).

[FIG] token + stem freq ggplot

title2_text <- glue::glue("Most frequent TOKEN & STEM in {n_distinct(pdo_train_t$proj_id)} PDOs") 
subtitle2_text <- glue::glue("From projects approved between FY {min(pdo_train_t$FY_appr)}-{max(pdo_train_t$FY_appr)}") 

combo_freq <-  pdo_wrd_freq_p + pdo_stem_freq_p + 
   plot_annotation(title = title2_text,
                    subtitle = subtitle2_text,
                   # caption = "Source: World Bank Project Documents",
                   theme = theme(plot.title = element_text(size = 12, face = "bold"),
                                 plot.subtitle = element_text(size = 10, face = "italic"),
                                 plot.caption = element_text(size = 10, face = "italic"))
                   )
combo_freq

f_save_plot("combo_freq", combo_freq)

_______

===== !!!!! QUI !!!!! ======

Check (by sector) why peaks [CMPL 🟠]

For the (broadly defined) health sector, it is quite clear that Covid-19 is the main driver of the peak in 2020. What about the other sectors? (they all seem to have at least 1 evident peak)

— I could see if corresponds to any WDR publications

Qualify: peak or trend (by sector) [CMPL 🟠]

… capire come si stabilisce che c’e’ un trend … magari vedere google search trend…

_______

BIGRAMS

_______

Here I use [clnp_annotate() output + ] dplyr to combine consecutive tokens into bigrams.

# Create bigrams by pairing consecutive tokens by sentence ID and token IDs
bigrams <- pdo_train_t %>%
   # keeping FY with tokens
   group_by(FY_appr, proj_id, pdo, sid ) %>%
   arrange(tid) %>%
   # Using mutate() and lead(), we create bigrams from consecutive tokens 
   mutate(next_token = lead(token), 
          bigram = paste(token, next_token)) %>%
   # make bigram low case
   mutate(bigram = tolower(bigram)) %>%
   # only includes the rows where valid bigrams are formed
   filter(!is.na(next_token)) %>%
   ungroup() %>%
   arrange(FY_appr, proj_id, sid, tid) %>%
   select(FY_appr,proj_id, pdo,sid, tid, token, bigram) 
# most frequent bigrams 
count_bigram <- bigrams %>% 
   count(bigram, sort = TRUE)

Clean bigrams

The challenge is to clean but without separating consecutive words… so I do this split-reunite process to remove stopwords and punctuation. Basically only keeping bigrams made of 2 nouns or ADJ+noun.

# Separate the bigram column into two words
bigrams_cleaned <- bigrams %>%
  tidyr::separate(bigram, into = c("word1", "word2"), sep = " ")

# Remove stopwords and bigrams in EACH component word containing punctuation
bigrams_cleaned <- bigrams_cleaned %>%
   # custom stop words
   filter(!word1 %in% custom_stop_words_df$word, !word2 %in% custom_stop_words_df$word) %>% 
   # Remove punctuation   
   filter(!str_detect(word1, "[[:punct:]]"), !str_detect(word2, "[[:punct:]]"))  

# Reunite the component cleaned words into the bigram column
bigrams_cleaned <- bigrams_cleaned %>%
   unite(bigram, word1, word2, sep = " ") %>% 
   # Remove too obvious bigrams 
   filter(!bigram %in% c("development objective", "development objectives", 
                         "proposed project", "project development", "program development"))

# View the cleaned dataframe
bigrams_cleaned

# Count the frequency of each bigram
bigram_freq <- bigrams_cleaned %>%
  count(bigram, sort = TRUE)

[FIG] most frequent bigrams in PDOs

  • Excluding bigrams where 1 word is among stopwords or a punctuation sign
  • Excluding “development objective/s”, “proposed project”, “program development” because not very informative
# ---- Prepare data for plotting
# Evaluate the title with glue first
title_text <- glue::glue("Frequency of bigrams in PDOs over FY {min(pdo_train_t$FY_appr)}-{max(pdo_train_t$FY_appr)}") 
# Define the bigrams you want to highlight
bigrams_to_highlight <- c("public sector", "private sector", "eligible crisis",
                          "health care", "health services", "public health")   

 
# ---- Plot the most frequent bigrams
pdo_bigr_freq <- bigram_freq %>%
   slice_max(n, n = 25) %>%
   ggplot(aes(x = reorder(bigram, n), y = n,
              fill = ifelse(bigram %in% bigrams_to_highlight, bigram, "Other"))) +
   geom_col() +
   # coord flipped so n is Y axis
   scale_y_continuous(breaks = seq(min(bigram_freq$n)-1, max(bigram_freq$n), by = 50)) +
   scale_fill_manual(values = c("public sector" = "#005ca1", 
                                "private sector" = "#9b2339", 
                                "eligible crisis"= "#8e550a", 
                                "health care"= "#245048",
                                "health services"= "#245048",
                                "public health"= "#245048", 
                                "Other" = "grey")) +
   guides(fill = "none") +
   coord_flip() +
   labs(title = title_text, subtitle = "(ranking top 25 bigrams)",
        x = "", y = "") +
   theme(axis.text.y = element_text(
            # obtain vector of colors 2 match x axis labels color to fill
            color = bigram_freq %>%
               slice_max(n, n = 25) %>%
               # mutate(color = ifelse(bigram %in% bigrams_to_highlight,
               #                       ifelse(bigram == "public sector", "#005ca1",
               #                              ifelse(bigram == "private sector", "#9b2339", "#8e550a")),
               #                       "#4c4c4c")) 
               mutate(color = case_when (
                  bigram == "public sector" ~ "#005ca1",
                  bigram == "private sector" ~ "#9b2339",
                  bigram == "eligible crisis" ~ "#8e550a",
                  bigram %in% c("health care", "health services", "public health") ~ "#245048",
                  TRUE ~ "#4c4c4c")) %>%
               # Ensure the order matches the reordered bigrams (AS BINS)
               arrange(reorder(bigram, n)) %>%  
               # Extract the color column in bin order as vector to be passed to element_text()
               pull(color)
            )
         ) + my_pretty_theme

pdo_bigr_freq

Results are not surprising in terms of frequent bigram recurrence:

  • See for example “increase access”, “service delivery” ,“institutional capacity”, “poverty reduction” etc, at the top.
  • Although, while “health” recurred in several bigrams (e.g. “health services”, “public health”, “health care”) among the top 25, “education” did not appear at all.
  • A bit mysterious is perhaps “eligible crisis” (> 100 mentions)?! (coming back to this later)

[FIG] Changes over time BY 1FY

Besides huge, counter intuitive, difference between “health” and “education”, “climate change” appears in the top 25 (ranking above “financial sector” and “capacity building”) which begs the question: Has the frequency of these bigrams has changed over time?

## too busy to be useful

# Step 1: Count the frequency of each bigram by year
top_bigrams_1FY <- bigrams_cleaned %>%
   group_by(FY_appr, bigram) %>%
   summarise(count = n(), .groups = 'drop') %>%
   arrange(FY_appr, desc(count)) %>%
   # ---  +/- top 10  
   group_by(FY_appr) %>%
   top_n(10, count) %>%
   ungroup()
   # # ---  STRICT  top 10  
   # mutate(rank = dense_rank(desc(count))) %>%  # Rank bigrams by frequency
   # filter(rank <= 10) %>%  # Keep only the top 10 by rank
   # ungroup()

  
# Add specific bigrams to highlight, if any
bigrams_to_highlight <- c("climate change",  "climate resilience", "public sector", "private sector")

# Step 2: Plot the top bigrams by frequency over time   
pdo_bigr_FY_freq  <-  top_bigrams_1FY %>% 
 ggplot(aes(x = reorder(bigram, count), 
             y = count,
             fill = ifelse(bigram %in% bigrams_to_highlight, bigram, "Other"))) +
  geom_col() +
  scale_fill_manual(values = c("public sector" = "#005ca1", "private sector" = "#e60066", 
                               "climate change" = "#399B23", "climate resilience" = "#d8e600",
                               "Other" = "grey")) +
  guides(fill = "none") +
  coord_flip() +
  facet_wrap(~ FY_appr, scales = "free_y") +
  labs(title = "Top 10 Bigrams by Frequency Over Time",
       subtitle = "(Faceted by Fiscal Year Approval)",
       x = "Bigrams",
       y = "Count") +
  theme_minimal() +
  theme(plot.title.position = "plot",
        axis.text.x = element_text(angle = 45, hjust = 1))+
     my_pretty_theme

pdo_bigr_FY_freq

[FIG] Changes over time BY 3FY

To reduce the noise and make the plot more readable, we can group the data by 3 fiscal years (FY) intervals.

# generate FY group 
f_generate_year_groups <- function(years, interval) {
  breaks <- seq(floor(min(years, na.rm = TRUE) / interval) * interval, 
                ceiling(max(years, na.rm = TRUE) / interval) * interval, 
                by = interval)
  
  labels <- paste(breaks[-length(breaks)], "-", breaks[-1] - 1)
  
  return(list(breaks = breaks, labels = labels))
}
# --- Step 1: Create n-year groups (using `f_generate_year_groups`)
interval_i = 3 # decide the interval
year_groups <- f_generate_year_groups(bigrams_cleaned$FY_appr, interval = interval_i)
top_n_i = 12 # decide the top n bigrams to show

# --- Step 2: Add the generated FY breaks and labels to data frame
top_bigrams_FYper <- bigrams_cleaned %>%
   # cut divides the range of x into intervals
   mutate(FY_group = base::cut(FY_appr, 
                               breaks = year_groups$breaks, 
                               include.lowest = TRUE, 
                               right = FALSE, 
                               labels = year_groups$labels)) %>% 
   # Count the frequency of each bigram by n-year groups
   group_by(FY_group, bigram) %>%
   summarise(count = n(), .groups = 'drop') %>%
   arrange(FY_group, desc(count)) %>%
   # Top ? bigrams for each n-year period
   group_by(FY_group) %>%
   top_n(top_n_i, count) %>%
   ungroup()

# --- Step 3: Add specific bigrams to highlight, if any
bigrams_to_highlight <- c("climate change",  "climate resilience", 
                          "eligible crisis",  
                          "public sector", "private sector",
                          "water supply", "sanitation services",
                          "health care", "health services", "public health", "health preparedness"
                          )

# --- Step 4: Plot the top bigrams by frequency over n-year periods
pdo_bigr_FY_freq  <-  top_bigrams_FYper %>% 
 ggplot(aes(x = reorder(bigram, count), 
             y = count,
             fill = ifelse(bigram %in% bigrams_to_highlight, bigram, "Other"))) +
  geom_col() +
  scale_fill_manual(values = c(
     # "public sector" = "#005ca1", 
     # "private sector" = "#e60066", 
     "water supply" = "#26BDE2",
      "sanitation services" = "#26BDE2",
     "climate change" = "#3F7E44", 
     "climate resilience" = "#a6bd23",
     "eligible crisis" = "#e68000",  
     "health care" = "#E5243B",
     "health services" = "#E5243B",
     "public health" = "#E5243B",
     "Other" = "grey")) +
  guides(fill = "none") +
  coord_flip() +
  facet_wrap(~ FY_group, ncol = 2 , scales = "free_y" )+ 
              #strip.position = "top") +  # Facet wrap with columns
  labs(title = glue::glue("Top 12 Bigrams by Frequency Over {interval_i}-Year Periods"),
       subtitle =  "(Some sectors highlighted)",
       x = "",
       y = "") +
     my_pretty_theme
# print the plot
pdo_bigr_FY_freq

Frequency observed over FY intervals is very revealing.

  • Interesting to see the trend of “water supply” and “sanitation services” bigrams, which are quite stable over time.
  • The bigram “health care” and “health services” are also quite stable, while “public health” obviously gained relevance since the 2019-2021 FY period.
  • Conversely, “private sector” and “public sector” loose importance over time (around mid 2010s), while “climate change” and “climate resilience” gain relevance from the same point on.
  • Still quite surprising the bigram “eligible crisis”, which actually appears in the top 12 bigrams starting in FY 2016-2018!

🤔 Which are the most frequent and persistent Bigrams Over Time?

For this, I am looking for a ranking that considers Mean frequency across periods arrange(desc(mean_count)) + Stability (low standard deviation) across periods [this is hard bc of NAs], and NOT total count overall…

  • Using top_bigrams_FYper which had breaks of 3FY
# ------------------------------[REPEATED just to see the table]

# --- Step 1: Create n-year groups (using `f_generate_year_groups`)
interval_i = 3 # decide the interval
year_groups <- f_generate_year_groups(bigrams_cleaned$FY_appr, interval = interval_i)
top_n_i = 12 # decide the top n bigrams to show

# --- Step 2: Add the generated FY breaks and labels to data frame
top_bigrams_FYper <- bigrams_cleaned %>%
   # cut divides the range of x into intervals
   mutate(FY_group = base::cut(FY_appr, 
                               breaks = year_groups$breaks, 
                               include.lowest = TRUE, 
                               right = FALSE, 
                               labels = year_groups$labels)) %>% 
   # Count the frequency of each bigram by n-year groups
   group_by(FY_group, bigram) %>%
   summarise(count = n(), .groups = 'drop') %>%
   arrange(FY_group, desc(count)) %>%
   # Top ? bigrams for each n-year period
   group_by(FY_group) %>%
   top_n(top_n_i, count) %>%
   ungroup()

sd() returns NA for bigrams that are not present in any periods (or are present in just 1 period).

# Calculate the mean frequency and standard deviation of the counts for each bigram across periods
stable_and_frequent_bigrams_per <- top_bigrams_FYper %>%
   group_by(bigram) %>%
   summarise(mean_count = mean(count, na.rm = TRUE),     # Mean frequency across periods
             sd_count = sd(count, na.rm = TRUE),         # Stability (lower sd = more stable)
             count_non_na = sum(!is.na(count)),  # Count non-NA values
             sd_count2 = if_else(count_non_na >= 1, sd(count, na.rm = TRUE), NA_real_),  # Only calculate sd if >= 3 non-NA
             total_count = sum(count)) %>%               # Total count across all periods (optional)
   arrange(desc(mean_count)) %>%                      # Sort by frequency and then stability
   # Filter out bigrams with low mean frequency or high instability (you can adjust thresholds)
   # Focus on the top 25% most frequent bigrams
   filter(mean_count > quantile(mean_count, 0.70, na.rm = TRUE)) #%>% 
   # Focus on the most stable 50% (lower sd) ---> NO bc NA values
   #filter( sd_count < quantile(sd_count, 0.5, na.rm = TRUE))

[TBL] Bigrams Over Time [3FY]

# View the most frequent and stable bigrams
stable_and_frequent_bigrams_per %>% 
   slice_head(n = 15)  %>% kableExtra::kable()
bigram mean_count sd_count count_non_na sd_count2 total_count
increase access 39.83333 6.080022 6 6.080022 239
eligible crisis 37.33333 1.527525 3 1.527525 112
threat posed 33.00000 NA 1 NA 33
private sector 31.20000 10.917875 5 10.917875 156
health preparedness 31.00000 NA 1 NA 31
strengthen national 28.00000 NA 1 NA 28
service delivery 27.71429 5.313953 7 5.313953 194
climate change 27.00000 2.828427 2 2.828427 54
poverty reduction 27.00000 14.514361 4 14.514361 108
public health 25.50000 16.263456 2 16.263456 51
public sector 25.25000 8.301606 4 8.301606 101
institutional capacity 24.87500 6.577831 8 6.577831 199
improve access 24.57143 8.521681 7 8.521681 172
national systems 24.00000 NA 1 NA 24
  • Using top_bigrams_1FY which had breaks of 1FY
# --- Step 1: Create n-year groups (using `f_generate_year_groups`)
interval_i = 1 # decide the interval
year_groups <- f_generate_year_groups(bigrams_cleaned$FY_appr, interval = interval_i)
top_n_i = 12 # decide the top n bigrams to show

# --- Step 2: Add the generated FY breaks and labels to data frame
top_bigrams_1FY <- bigrams_cleaned %>%
   # cut divides the range of x into intervals
   mutate(FY_group = base::cut(FY_appr, 
                               breaks = year_groups$breaks, 
                               include.lowest = TRUE, 
                               right = FALSE, 
                               labels = year_groups$labels)) %>% 
   # Count the frequency of each bigram by n-year groups
   group_by(FY_group, bigram) %>%
   summarise(count = n(), .groups = 'drop') %>%
   arrange(FY_group, desc(count)) %>%
   # Top ? bigrams for each n-year period
   group_by(FY_group) %>%
   top_n(top_n_i, count) %>%
   ungroup()
# Calculate the mean frequency and standard deviation of the counts for each bigram across periods
stable_and_frequent_bigrams_1FY <- top_bigrams_1FY %>%
   group_by( bigram) %>%
   summarise(mean_count = mean(count, na.rm = TRUE),     # Mean frequency across periods
             sd_count = sd(count, na.rm = TRUE),         # Stability (lower sd = more stable)
             total_count = sum(count)) %>%               # Total count across all periods (optional)
   arrange(desc(mean_count)) %>%                      # Sort by frequency and then stability
   # Filter out bigrams with low mean frequency or high instability (you can adjust thresholds)
   # Focus on the top 25% most frequent bigrams
   filter(mean_count > quantile(mean_count, 0.70, na.rm = TRUE)) #%>% 
   # Focus on the most stable 50% (lower sd) ---> NO bc NA values
   #filter( sd_count < quantile(sd_count, 0.5, na.rm = TRUE))

[TBL] Bigrams Over Time [1FY]

# View the most frequent and stable bigrams
stable_and_frequent_bigrams_1FY %>% 
   slice_head(n = 15)   %>% kableExtra::kable()
bigram mean_count sd_count total_count
mobile applications 21.00000 NA 21
public health 16.66667 3.0550505 50
threat posed 16.50000 2.1213203 33
health preparedness 15.50000 0.7071068 31
increase access 14.64706 5.1713293 249
eligible crisis 14.62500 10.1971635 117
strengthen national 14.00000 2.8284271 28
vulnerable households 13.00000 NA 13
respond promptly 12.50000 10.6066017 25
action plan 12.00000 NA 12
disaster risk 12.00000 NA 12
local governments 12.00000 NA 12
national systems 12.00000 1.4142136 24
world bank 12.00000 NA 12
climate resilience 11.66667 4.5092498 35

_______

Explore specific bigrams

— Public/Private ~ compare frequency over FY

A case in which looking at bigrams may be better than tokens is the question whether WB project are more focused on public or private sector. It is not easy to capture this information from the text, because:

  • “government” may be referred to the subject/counterpart of the project (e.g. “government of Mozambique”)
  • “private” is not necessarily referred to the “private sector” (e.g. “private households”)
  • “public” is not necessarily referred to the “public sector” (e.g. “public health”)

So, I narrow down to consecutive bigrams “public sector” and “private sector” to get an indicative frequency of these terms.

[FIG] Bigrams (“public sector”, “private sector”) freq plot

# Filter for the specific bigrams "public sector" and "private sector"
bigrams_pub_priv_sec <- bigrams %>%
   filter(bigram %in% c("public sector", "private sector"))

# Display the result
#bigrams_pub_priv_sec

# prepare data for plotting (count)
sector_bigr_df <- bigrams_pub_priv_sec %>% 
   count(FY_appr, bigram) %>% 
   # reorder values by frequency
   mutate(bigram = factor(bigram, levels = c("public sector", "private sector")))
# ---- Prepare data for plotting
# Evaluate the title with glue first
title_text <- glue::glue("Frequency of bigrams \"public sector\" and \"private sector\" in PDOs over FY {min(sector_bigr_df$FY_appr)}-{max(sector_bigr_df$FY_appr)}") 

two_col_contrast <- c( "#005ca1",  "#e60066" )

# Create a named vector for the legend labels with totals in a single pipeline
legend_labels <- sector_bigr_df %>%
   group_by(bigram) %>%
   # Calculate total counts for each bigram
   summarize(total_n = sum(n)) %>% 
   # Append totals to bigram names
   mutate(label = paste0(bigram, " (", total_n, ")")) %>%  
   # Create a named vector with bigram as names and labels as values
   {setNames(.$label, .$bigram)} # curly braces {} in a dplyr pipeline using . as ouptu from previous..

# ---- Plot
pdo_pub_pri_bigr <- ggplot(data = sector_bigr_df, aes(x = FY_appr, y = n, group = bigram, color = bigram)) +
   geom_line(linetype = "dotted", alpha = 0.75, size = 1) +
   geom_point(size = 3) +
   scale_x_continuous(breaks = seq(2001, 2023, by = 1)) +
   scale_color_manual(values = two_col_contrast, 
                      labels = legend_labels) +  # Use modified labels
   my_pretty_theme +
   theme(axis.text.x = element_text(angle = 45, hjust = 1)) +
   labs(title = title_text, 
        x = "", 
        y = "", 
        color = "") 
   

pdo_pub_pri_bigr

# Save the plot
f_save_plot("pdo_pub_pri_bigr", pdo_pub_pri_bigr)

Note:

  • these are much less common than the single words.
  • What happens in FY 2014-2016 that makes these bigram drop in frequency of mention?

— climate change ~ notable bigrams over FY [CMPL 🟠]

— eligible crisis ~ notable bigrams over FY

# reduce back to the original data
pdo_t <- pdo_train_t %>% 
   select(proj_id, pdo,pr_name, FY_appr, FY_clos, status, regionname, countryname, sector1, theme1,lendinginstr, env_cat, ESrisk, curr_total_commitment) %>%
   group_by(proj_id) %>% 
   slice(1)

First of all, let’s see what are the sentence that contain the bigram “eligible crisis” in the PDOs.

# ---- Define the bigram you want to find
target_bigram <- "eligible crisis"

# Tokenize the text data into sentences
sentences <- pdo_t %>%
   unnest_tokens(sentence, pdo, token = "sentences")

# Count the number of sentences in each document
sentence_count <- sentences %>%
   group_by(proj_id) %>%
   summarise(num_sentences = n())

n_distinct(sentence_count$proj_id)  # number of projects
sum(sentence_count$num_sentences)   # total number of sentences

# Filter sentences that contain the specific bigram
sentences_with_targ <- sentences %>%
   filter(str_detect(sentence, target_bigram))

# Define how many characters before and after the bigram to extract
chars_before <- 60  # Number of characters before the bigram
chars_after <- 60   # Number of characters after the bigram

# Add the extracted bigram and surrounding characters to the same dataframe
sentences_with_eligcris <- sentences_with_targ %>%
   mutate(closest_text = str_extract(sentence, paste0(".{0,", chars_before, "}", target_bigram, ".{0,", chars_after, "}"))) %>% 
   # View the updated dataframe with the closest_text column
   select(proj_id, #sentence, 
          closest_text)

[TBL] Close phrase around bigram “eligible crisis”

I still don’t know what “eligible” crisis means, but it appears that the following is like a commonly used phrase in the PDOs: “to respond promptly and effectively” as well as “provide immediate and effective response to” seem to often accompany the text “eligible crisis or emergency”. Presumably, a standard sentence indicating eligibility for ODA funding.

# Define the phrase you want to search for in the vicinity of the target bigram
phrase_to_search <- "respond promptly and effectively"

# Count how often the phrase appears in the vicinity of the target bigram
phrase_count <- sentences_with_eligcris %>%
  mutate(contains_phrase = str_detect(closest_text, phrase_to_search)) %>%  # Check if the phrase is present
  summarise(count = sum(contains_phrase))  # Count how many times the phrase is found

# View the result
tabyl(phrase_count$count)

Here are a few examples of the sentences containing the bigram “eligible crisis” and the phrase “respond promptly and effectively”:

# Filter the sentences that contain the phrase
sample_with_eligcris <-  sentences_with_eligcris %>% 
   ungroup() %>% 
   # take a random sample of 5 sentences
   sample_n(8 ) %>%
   select(proj_id, closest_text) %>% 
   mutate (closest_text =  paste0("(...) ", closest_text),
           # Make "eligible crisis" bold by adding <b> tags
           closest_text = gsub("eligible crisis", "<b>eligible crisis</b>", closest_text)
   )

# print out sample in a kable 
kable(sample_with_eligcris, format = "html", 
      # Display the table with bold formatting
       escape = FALSE,
      col.names = c("WB Project ID","excerpt of PDO sentences with 'eligible crisis'")) %>% 
   kable_styling(full_width = FALSE)
WB Project ID excerpt of PDO sentences with 'eligible crisis'
P172504 (...) reas, and to provide immediate and effective response to an eligible crisis or emergency
P176304 (...) provide immediate and effective response in the case of an eligible crisis or emergency
P164431 (...) rigation and provide immediate and effective response to an eligible crisis or emergency.
P175587 (...) sme activities in the batken region; and (iv) in case of an eligible crisis or emergency, respond promptly and effectively to it.
P127338 (...) rrower's capacity to respond promptly and effectively in an eligible crisis or emergency, asrequired.
P177100 (...) and safety of maritime sector in tuvalu, and in case of an eligible crisis or emergency, respond promptly and effectively to it.
P176905 (...) gri-food system; and (ii) respond effectively in case of an eligible crisis or emergency.
P175172 (...) ions, and to provide immediate and effective response to an eligible crisis or emergency.

Analyzing bigrams: tf-idf

There are advantages and disadvantages to examining the tf-idf of bigrams rather than individual words. Pairs of consecutive words might capture structure that isn’t present when one is just counting single words, and may provide context that makes tokens more understandable. However, the per-bigram counts are also sparser (a typical two-word pair is rarer than either of its component words).

bigram_tf_idf <- bigrams_cleaned %>% 
 # then on that calculate tf-idf
 count(FY_appr, bigram) %>%
  bind_tf_idf(bigram, FY_appr, n) %>%
  arrange(FY_appr, desc(tf_idf))

bigram_tf_idf_top <- slice_head(bigram_tf_idf, n =  5, by = FY_appr ) %>% 
      arrange(FY_appr, desc(tf_idf))

_______

Acknowledgements

Below are some invaluable resources that I used to learn and implement the NLP techniques in this analysis: + An Introduction to Quantitative Text Analysis for Linguistics